
# This analysis aims to explore and segment customer behavior patterns using a bunch of unsupervised learning algorithms.
# The goal is to identify meaningful groups of customers based on their financial activity and credit behavior, enabling 
# deeper understanding of customer types and potential detection of their spending patterns.
# The dataset includes several numerical variables related to customer credit usage, such as balance, purchases, payments,
# credit limit, cash advances, and payment frequency. Before applying any clustering model,some data preprocessing and EDA
# will be conducted to ensure data quality and comparability. This includes handling missing values, treating outliers, 
# transforming skewed distributions using log and Box-Cox transformations, and scaling all numerical features for 
# uniformity.
# After preprocessing, multiple clustering techniques will be applied and compared to identify the most suitable approach for
# this dataset.


import pandas as pd
import numpy as np   
df = pd.read_csv('Customer_Data.csv')
df


df.dtypes

CUST_ID                              object
BALANCE                             float64
BALANCE_FREQUENCY                   float64
PURCHASES                           float64
ONEOFF_PURCHASES                    float64
INSTALLMENTS_PURCHASES              float64
CASH_ADVANCE                        float64
PURCHASES_FREQUENCY                 float64
ONEOFF_PURCHASES_FREQUENCY          float64
PURCHASES_INSTALLMENTS_FREQUENCY    float64
CASH_ADVANCE_FREQUENCY              float64
CASH_ADVANCE_TRX                      int64
PURCHASES_TRX                         int64
CREDIT_LIMIT                        float64
PAYMENTS                            float64
MINIMUM_PAYMENTS                    float64
PRC_FULL_PAYMENT                    float64
TENURE                                int64
dtype: object


# columns


# 1. CUST_ID: index for the custumers
    
# 2. BALANCE: The avg amount of money that every custmers owed which is shown on their credit card. High amount is 
# showing the custimers had high debt and low balance can show the coustomers pay off the bills often or use the cards
# less.

# 3.BALANCE_FREQUENCY: Showing the frequncy of the cars updated or use. Whether it is active or not active. the value is 
# between 0 and 1 which 1 means the card is highly active. 

# 4. PURCHASES: The total value of all purchases made by the customer during the observation period. If it is high, it 
# means the customer is heavy spender. In other word it is about spending level. 

# 5.ONEOFF_PURCHASES: Amount of purchases made in a single payment, not in installments. It is a kind of shwoing the
# payment inference of the custimers and if the value is high then  customer prefers one-time payments or possibly 
# short-term expenses otherwise customer prefers installment plans.

# 6. INSTALLMENTS_PURCHASES: Total amount of purchases made in installments. if it is high, customer prefers paying in
# parts over time  and they are budget-conscious or large purchases and low value showing rarely uses installment options.
# it is the indication of the financing behavior.

# 7. CASH_ADVANCE: Total amount of cash withdrawn using the credit card. it can be like taking a short term loan.The high
# value is the indicator of customer have liquidity issues and low value shoing the customet most o fthe time using the 
# card for purchase. It is a good indicator for Good indicator of financial stress or emergency borrowing.

# 8. PURCHASES_FREQUENCY: Measures how often the customer uses the card for shopping. it indicates the sepnding regulari
# and its value is between 0 and 1 in which 1 is the active spender. 

# 9. ONEOFF_PURCHASES_FREQUENCY: It is shwoing if customer often makes single-payment purchases (1 ) or never or 
# rarely does ( 0).

# 10. PURCHASES_INSTALLMENTS_FREQUENCY: It indicates credit usage style. if it is high, the customer frequently pays in 
# installments and low mean prefers full payments.

# 11. CASH_ADVANCE_FREQUENCY: Frequency (0–1) with which the customer takes cash advances. If it is high, it will show 
# the customers relies more on credit for cash and low means they rarly relu on cash.

# 12. CASH_ADVANCE_TRX:Showing how often the person takes out cash. 
# 13. PURCHASES_TRX: Total number of purchase transactions and shows transaction behavior intensity. If it is high,it wil
# show the customers frequently use card  for purchases.

# 14. CREDIT_LIMIT: The maximum credit limit allowed for the customer. id the value is high for this column , it means
# customer has good credit history or income and low value means limited trust or new user. It is Indicator of customer
# risk level or creditworthiness.

# 15. PAYMENTS:Total amount of money paid back by the customer. If it is high, it means the costumers are 
# financially responsible and pays frequently and low means may miss or delay payments.

# 16. MINIMUM_PAYMENTS: The smallest amount a customer must pay on their credit card each month to avoid a late fee. Pay
# off their balance fully or mostly mean low risk), or barely meet the minimum payment  means potentially risky.
# payments, customer pays more than required.If close to total, the customers only pays the minimum (riskier behavior).

# 17. PRC_FULL_PAYMENT: Proportion of months in which the full balance was paid off (0–1).1 = pays full amount every month
# 0 = never pays full amount. It is a good measure of financial discipline.

# 18. TENURE: Number of months the customer has been active (usually 12 in this dataset).Longer tenure  equals 
# loyal customer, shorter means new customer. It is usful for loyalty and retention analysis.


# Spending behavior: The follwoing columns show how much and how often people buy:

# PURCHASES
# ONEOFF_PURCHASES
# INSTALLMENTS_PURCHASES
# PURCHASES_FREQUENCY
# ONEOFF_PURCHASES_FREQUENCY
# PURCHASES_INSTALLMENTS_FREQUENCY
# PURCHASES_TRX


# Credit and debt:The follwoing columns show how much debt customers carry or are allowed:

# BALANCE
# BALANCE_FREQUENCY
# CREDIT_LIMIT


# Cash usage: The following columns show if customers use their card to withdraw cash:

# CASH_ADVANCE
# CASH_ADVANCE_TRX
# CASH_ADVANCE_FREQUENCY


# Payment behavior: The follwoingcoumns show how well customers repay:

# PAYMENTS
# MINIMUM_PAYMENTS
# PRC_FULL_PAYMENT


# Customer profile: 

# TENURE : how long the customer has been active
# CUST_ID : just an identifier (not used in clustering)


df.describe()


len(df)

8950


import matplotlib.pyplot as plt

plt.figure(figsize=(6,4))
plt.scatter(df["PURCHASES"], df["PAYMENTS"], alpha=0.6)
plt.xlabel("PURCHASES (Total Spending)")
plt.ylabel("PAYMENTS (Total Repayments)")
plt.title("Spending vs. Repayment")
plt.show()


# This plot displays how customers behave in spending and repayment. Those in the bottom-left spend and pay little, likely 
# inactive or new users. The top-right group spends and pays a lot, showing healthy, active usage which are very few.
# Points far right but low are heavy spenders who don’t repay much and are potentially risky, while top-left customers pay
# more than they spend, likely settling previous debts.


plt.figure(figsize=(6,4))
plt.scatter(df["CASH_ADVANCE"], df["PURCHASES"], alpha=0.6, color='orange')
plt.xlabel("CASH_ADVANCE (Borrowed Cash)")
plt.ylabel("PURCHASES (Total Spending)")
plt.title("Borrowing vs. Spending")
plt.show()


# This plot shows the relationship between borrowing and spending. Most customers have low cash advances and low purchase
# amounts, meaning they rarely use their card to borrow cash. A few on the right borrow large amounts but spend little,
# suggesting they mainly use credit cards for emergency cash rather than shopping. Overall, high borrowing with low 
# spending may signal financial stress or liquidity problems.


plt.figure(figsize=(6,4))
plt.scatter(df["CREDIT_LIMIT"], df["BALANCE"], alpha=0.6, color='green')
plt.xlabel("CREDIT_LIMIT")
plt.ylabel("BALANCE (Outstanding Amount)")
plt.title("Credit Capacity vs. Usage")
plt.show()


# This plot shows how customers use their available credit. Most points form an upward triangle, meaning higher credit 
# limits often come with higher balances, but not always full utilization. Customers near the top edge use much of their
# limit and may be at higher financial risk, while those far below the diagonal use only a small part of their credit, 
# showing safer, more controlled spending.


plt.figure(figsize=(6,4))
plt.scatter(df["PURCHASES_FREQUENCY"], df["PRC_FULL_PAYMENT"], alpha=0.6, color='purple')
plt.xlabel("PURCHASES_FREQUENCY")
plt.ylabel("PRC_FULL_PAYMENT (Full Payment Ratio)")
plt.title("Usage Frequency vs. Responsibility")
plt.show()


# This plot shows how often customers make purchases compared to how responsibly they pay their bills. Many customers have
# high purchase frequency but low full-payment ratios, meaning they use their cards a lot but rarely clear their balance. 
# Only a few combine frequent use with full repayment, showing strong financial discipline. Overall, it highlights 
# different spending habits and responsibility levels across users.


plt.figure(figsize=(6,4))
plt.scatter(df["ONEOFF_PURCHASES"], df["INSTALLMENTS_PURCHASES"], alpha=0.6, color='red')
plt.xlabel("ONEOFF_PURCHASES (One-time)")
plt.ylabel("INSTALLMENTS_PURCHASES")
plt.title("Payment Method Preference")
plt.show()


# This plot compares customers preference for one-time versus installment purchases. Most points cluster near the 
# bottom-left, meaning most customers make small one-time purchases and rarely use installment plans.Most customers make 
# small one-time payments and rarely buy in installments A few toward the
# right rely more on one-off payments, while some higher up use installments more often, possibly for larger or planned
# purchases. It reveals each customer’s preferred payment style and credit usage pattern.

# The point far left (close to the x-axis origin, around 0 on ONEOFF_PURCHASES) means that customer didn’t make any 
# one-time purchases, but they might still have installment purchases.
# The point far above (very high INSTALLMENTS_PURCHASES value) represents a customer who spent a large amount using
# installments — likely a heavy buyer who prefers paying over time.

# It means They didn’t buy anything they paid for all at once, but they made a big purchase that they chose to pay in
# parts over time.They only use their credit card for big, planned purchases paid gradually, not for everyday shopping.

# Notice: the value in INSTALLMENTS_PURCHASES is the sum of all installment purchases a person made during the entire
# period in the dataset, not just one single item or plan.
# The number shows the total dollar amount of all those installment-based purchases combined.


df.isnull().sum()

CUST_ID                               0
BALANCE                               0
BALANCE_FREQUENCY                     0
PURCHASES                             0
ONEOFF_PURCHASES                      0
INSTALLMENTS_PURCHASES                0
CASH_ADVANCE                          0
PURCHASES_FREQUENCY                   0
ONEOFF_PURCHASES_FREQUENCY            0
PURCHASES_INSTALLMENTS_FREQUENCY      0
CASH_ADVANCE_FREQUENCY                0
CASH_ADVANCE_TRX                      0
PURCHASES_TRX                         0
CREDIT_LIMIT                          1
PAYMENTS                              0
MINIMUM_PAYMENTS                    313
PRC_FULL_PAYMENT                      0
TENURE                                0
dtype: int64


# filling the CREDIT_LIMIT with MOde 
mode_value = df['CREDIT_LIMIT'].mode()[0] 
df['CREDIT_LIMIT'] = df['CREDIT_LIMIT'].fillna(mode_value)

print(df['CREDIT_LIMIT'].isnull().sum())
print(mode_value)

0
3000.0


# The two methos of filling missing value for Min Payment would be evaluated: 1- Median 2: KNN

from sklearn.impute import KNNImputer

# Median 
df1 = df.copy()
median_value = df1['MINIMUM_PAYMENTS'].median()
df1['MINIMUM_PAYMENTS'] = df1['MINIMUM_PAYMENTS'].fillna(median_value)

# KNN
df2 = df.copy()

num_cols = df2.select_dtypes(include=['float64', 'int64']).columns

imputer = KNNImputer(n_neighbors=5)
df2[num_cols] = imputer.fit_transform(df2[num_cols])

print("Missing values in df1 (median method):")
print(df1['MINIMUM_PAYMENTS'].isnull().sum())

print("\nMissing values in df2 (KNN method):")
print(df2['MINIMUM_PAYMENTS'].isnull().sum())

Missing values in df1 (median method):
0

Missing values in df2 (KNN method):
0


# Visulizing both methods with original Min Payment to see which one perform better. 

import seaborn as sns
import matplotlib.pyplot as plt

original_non_missing = df['MINIMUM_PAYMENTS'].dropna()

plt.figure(figsize=(8,5))
sns.kdeplot(original_non_missing, label='Before Imputation (Original Non-Missing)', linewidth=2)
sns.kdeplot(df1['MINIMUM_PAYMENTS'], label='After Median Imputation', linestyle='--')
sns.kdeplot(df2['MINIMUM_PAYMENTS'], label='After KNN Imputation', linestyle=':')
plt.legend()
plt.title("Distribution Comparison of MINIMUM_PAYMENTS: Before vs After Imputation")
plt.xlabel("MINIMUM_PAYMENTS")
plt.ylabel("Density")
plt.show()


# Both imputation methods (median and Knn) distribution seems like the original one.
# both perform well. for simplicity, the median approch will be kept.


# Another method to check their performance is Corrolation Criteria 
# if both imputation methods preserve the same correlation patterns, then both are fine.
# If one changes correlations drastically, it might be distorting multi-variable relationships.
corr_before = df.corr(numeric_only=True)['MINIMUM_PAYMENTS']
corr_median = df1.corr(numeric_only=True)['MINIMUM_PAYMENTS']
corr_knn = df2.corr(numeric_only=True)['MINIMUM_PAYMENTS']

comparison = pd.DataFrame({
    'Before': corr_before,
    'After_Median': corr_median,
    'After_KNN': corr_knn
})

print(comparison)

                                    Before  After_Median  After_KNN
BALANCE                           0.398684      0.397920   0.400514
BALANCE_FREQUENCY                 0.132569      0.131181   0.137702
PURCHASES                         0.093860      0.095789   0.095084
ONEOFF_PURCHASES                  0.048755      0.050256   0.050064
INSTALLMENTS_PURCHASES            0.132172      0.134019   0.132704
CASH_ADVANCE                      0.140107      0.140747   0.141165
PURCHASES_FREQUENCY               0.003030      0.006154   0.005053
ONEOFF_PURCHASES_FREQUENCY       -0.030327     -0.027311  -0.027291
PURCHASES_INSTALLMENTS_FREQUENCY  0.030073      0.032055   0.030875
CASH_ADVANCE_FREQUENCY            0.098838      0.100616   0.102743
CASH_ADVANCE_TRX                  0.109739      0.111251   0.111752
PURCHASES_TRX                     0.096333      0.098722   0.097950
CREDIT_LIMIT                      0.126686      0.126765   0.125305
PAYMENTS                          0.126651      0.128960   0.126585
MINIMUM_PAYMENTS                  1.000000      1.000000   1.000000
PRC_FULL_PAYMENT                 -0.140379     -0.135194  -0.136101
TENURE                            0.059444      0.060006   0.060771


# it showing no difference and we go with median.
df1


# Trying visualization and see which variables are normally distributed , have outliers and also the realtionship betw
# different variables which are the most imporatnt section for unsupervised learning.


numeric_cols = df1.select_dtypes(include=['float64', 'int64']).columns

n_cols = 4
n_rows = (len(numeric_cols) + n_cols - 1) // n_cols 

plt.figure(figsize=(18, n_rows * 3)) 

for i, col in enumerate(numeric_cols, 1):
    plt.subplot(n_rows, n_cols, i)
    sns.histplot(df1[col], bins=30, kde=False, color='steelblue')
    plt.title(col, fontsize=10)
    plt.xlabel('')
    plt.ylabel('')

plt.tight_layout(pad=2.5) 
plt.suptitle("Distribution of Numeric Features (After Median Imputation)", fontsize=16, y=1.02)
plt.show()


numeric_cols = df1.select_dtypes(include=['float64', 'int64']).columns

rows, cols = 5, 4

plt.figure(figsize=(20, 18))

for i, col in enumerate(numeric_cols):
    plt.subplot(rows, cols, i + 1)
    sns.boxplot(x=df1[col], color='lightcoral')
    plt.title(f"{col}", fontsize=10)
    plt.xlabel('')
    plt.ylabel('')

plt.tight_layout(pad=2.5)
plt.suptitle("Boxplots of All Numeric Features (5x4 Layout)", fontsize=16, y=1.02)
plt.show()


# TENURE variable (number of months as a customer) is almost constant across rows likely all 12 or very close to it.
# Since the clustering works based on variation across features and distance calculation, we need to remove it for the
# sake of simplicity and not intriducing noise. 
# for the rest with continous (unbounded) features, log or cox box and sqrt trasformation would be applied to make them
# see normalized.


# pairwise corrolation matrix before transformation 

numeric_cols = df1.select_dtypes(include=['float64', 'int64']).columns

rows, cols = 5, 4

plt.figure(figsize=(20, 18))

for i, col in enumerate(numeric_cols):
    plt.subplot(rows, cols, i + 1)
    sns.kdeplot(df1[col], fill=True, color='steelblue', linewidth=1.2)
    plt.title(f"{col}", fontsize=10)
    plt.xlabel('')
    plt.ylabel('')

plt.tight_layout(pad=2.5)
plt.suptitle("KDE (Density) Plots of All Numeric Features (5x4 Layout)", fontsize=16, y=1.02)
plt.show()


# Heatmap corrolation

plt.figure(figsize=(12, 10))
sns.heatmap(
    df1[numeric_cols].corr(),
    cmap='coolwarm',
    annot=True,           
    fmt=".2f",            
    annot_kws={"size": 8},
    linewidths=0.5        
)
plt.title("Correlation Heatmap of Numeric Features", fontsize=14, pad=12)
plt.xticks(rotation=45, ha='right')
plt.yticks(rotation=0)
plt.tight_layout()
plt.show()


# Pair Plot (Feature Interactions)
sns.pairplot(df1[['BALANCE', 'PURCHASES', 'CASH_ADVANCE', 'PAYMENTS']], diag_kind='kde')
plt.suptitle("Pairwise Relationships Among Key Financial Variables", y=1.02)
plt.show()


# Removing TENURE column
df1 = df1.drop(columns=['TENURE'])


# square root transformation to transaction count features. this transformation method is used for the foloowing
# variables because their magnitudes are limited. 

df1['CASH_ADVANCE_TRX'] = np.sqrt(df1['CASH_ADVANCE_TRX'])
df1['PURCHASES_TRX'] = np.sqrt(df1['PURCHASES_TRX'])

df1[['CASH_ADVANCE_TRX', 'PURCHASES_TRX']].head()


# applying log trans

df3 = df1.copy()

log_features = ['BALANCE', 'PURCHASES', 'CASH_ADVANCE', 'PAYMENTS', 'MINIMUM_PAYMENTS']

df3[log_features] = np.log1p(df3[log_features])

df3[log_features].head()


# Applying Box-cCox

from scipy.stats import boxcox

df4 = df1.copy()

for col in ['BALANCE', 'PURCHASES', 'CASH_ADVANCE', 'PAYMENTS', 'MINIMUM_PAYMENTS']:
    df4[col], _ = boxcox(df4[col] + 1)

df4[['BALANCE', 'PURCHASES', 'CASH_ADVANCE', 'PAYMENTS', 'MINIMUM_PAYMENTS']].head()


# Evaluating 


features = ['BALANCE', 'PURCHASES', 'CASH_ADVANCE', 'PAYMENTS', 'MINIMUM_PAYMENTS']

comparison = pd.DataFrame({
    'Feature': features,
    'Skew_Before': df1[features].skew(),
    'Skew_Log': df3[features].skew(),
    'Skew_BoxCox': df4[features].skew(),
    'Kurt_Before': df1[features].kurt(),
    'Kurt_Log': df3[features].kurt(),
    'Kurt_BoxCox': df4[features].kurt()
}).round(2)

comparison


# Before transformation all features were strongly right-skewed (2–14 range) but after transformation Box–Cox achieved 
# near 0 skew (best for 4 of 5 features) and Log improved but remained slightly asymmetric.So Box–Cox performs better 
# overall for the data and we wiil go with it.

df4


# Now we go for scaling which is the most important part for clustering.


from sklearn.preprocessing import StandardScaler

df_scaled = df4.copy()

numeric_cols = df_scaled.select_dtypes(include=['float64', 'int64']).columns

scaler = StandardScaler()

df_scaled[numeric_cols] = scaler.fit_transform(df_scaled[numeric_cols])

df_scaled.describe().round(2)


df_scaled


customer_ids = df_scaled['CUST_ID']
df_cluster = df_scaled.drop(columns=['CUST_ID'])


corr_matrix = df_cluster.corr()

plt.figure(figsize=(12, 10))
sns.heatmap(
    corr_matrix,
    cmap='coolwarm',
    center=0,
    annot=True,       
    fmt=".2f",
    annot_kws={"size": 8}
)
plt.title("Correlation Heatmap After Scaling", fontsize=14, pad=12)
plt.xticks(rotation=45, ha='right')
plt.yticks(rotation=0)
plt.tight_layout()
plt.show()


# Since in clustering (especially K-Means or PCA), highly correlated features can overweight certain behavioral dimensions,
# reduce interpretability, and cause artificial clustering directions (since correlated variables move together).
# So we can remove one variable from each highly correlated pair (correlation > 0.85).


# Features with high unique value that are less correlated and should definitely stay:

# 1-BALANCE
# 2-BALANCE_FREQUENCY
# 3-ONEOFF_PURCHASES
# 4-INSTALLMENTS_PURCHASES
# 5-CREDIT_LIMIT
# 6-PRC_FULL_PAYMENT
# 7-PAYMENTS

# They each represent distinct behavioral or financial characteristics.


# Droping redundant features based on high correlation
df_final = df_scaled.drop(columns=[
    'PURCHASES', 
    'CASH_ADVANCE', 
    'PURCHASES_INSTALLMENTS_FREQUENCY',
    'CASH_ADVANCE_FREQUENCY'
])

df_final.shape

(8950, 13)


df_final


# removing CUST_ID ti have just meaningful numerical features for clustering. We will use it later on the cluster graphs
df_final_with_id = df_final.copy()
df_f = df_final_with_id.drop(columns=['CUST_ID'])
df_f


print(df_f.columns.tolist())

['BALANCE', 'BALANCE_FREQUENCY', 'ONEOFF_PURCHASES', 'INSTALLMENTS_PURCHASES', 'PURCHASES_FREQUENCY', 'ONEOFF_PURCHASES_FREQUENCY', 'CASH_ADVANCE_TRX', 'PURCHASES_TRX', 'CREDIT_LIMIT', 'PAYMENTS', 'MINIMUM_PAYMENTS', 'PRC_FULL_PAYMENT']


# Using K-mean cluster
# finding the best K with elbow method

from sklearn.cluster import KMeans

K = range(2, 11)

inertia = []  

for k in K:
    kmeans = KMeans(n_clusters=k, random_state=42, n_init=10)
    kmeans.fit(df_f)
    inertia.append(kmeans.inertia_)


plt.figure(figsize=(8, 5))
plt.plot(K, inertia, 'bo-')
plt.xlabel('Number of Clusters (k)')
plt.ylabel('Inertia (Within-Cluster Sum of Squares)')
plt.title('Elbow Method for Optimal k')
plt.grid(True)
plt.show()


# finding the best K with silhoutte method

from sklearn.metrics import silhouette_score

K = range(2, 11)

silhouette_scores = []

for k in K:
    kmeans = KMeans(n_clusters=k, random_state=42, n_init=10)
    kmeans.fit(df_f)
    score = silhouette_score(df_f, kmeans.labels_)
    silhouette_scores.append(score)

plt.figure(figsize=(8, 5))
plt.plot(K, silhouette_scores, 'go-')
plt.xlabel('Number of Clusters (k)')
plt.ylabel('Silhouette Score')
plt.title('Silhouette Score for Optimal k')
plt.grid(True)
plt.show()


kmeans = KMeans(n_clusters=4, random_state=42, n_init=10)
kmeans.fit(df_f)

df_final_with_id['Cluster'] = kmeans.labels_

df_final_with_id['Cluster'].value_counts()

Cluster
1    3487
3    2765
2    1444
0    1254
Name: count, dtype: int64


numeric_cols = df_final_with_id.select_dtypes(include=['number']).columns
cluster_summary = df_final_with_id.groupby('Cluster')[numeric_cols].mean().round(2)

cluster_summary['Count'] = df_final_with_id['Cluster'].value_counts()
cluster_summary = cluster_summary.sort_index()

cluster_summary


# In above table each column is a feature (behavior metric), and each row (Cluster 0–3) shows the average standardized 
# score for that feature. Positive  means above average and Negative means below average.
# because purcahse was highly related to other columns we removed it but we assume that :
# PURCHASES=ONEOFF_PURCHASES+INSTALLMENTS_PURCHASES

# Cluster 0 : High Spenders, active and Responsible Payers(1,254 customers): 
# Their value for PURCHASES, INSTALLMENTS_PURCHASES, PURCHASES_TRX , ONEOFF_PURCHASES_FREQUENCY are high which showing
# they are frequent and heavy on spending and spending a lot spend a lot both in single payments and installments and are 
# very active users because of PURCHASES_TRX and PURCHASES_FREQUENCY. 
# Their BALANCE and CREDIT_LIMIT showing that they are Above average, so they maintain larger balances but they can 
# control it. 
# Their payment behavior(PAYMENTS ,MINIMUM_PAYMENTS,PRC_FULL_PAYMEN  ) showing that they have Strong repayment behavior, 
# pay more than minimum often, and occasionally clear balances fully.
# So it can be said They are premium, high-activity ,high-value customers and heavy card users who also make regular, 
# strong payments.They are profitable and low-risk for the bank.


# Cluster 1 :Moderate Spenders with Cash Dependency (3,487 customers): 
# They Frequently use their card for cash withdrawals(CASH_ADVANCE_TRX = 0.70), Rarely use their card for shopping
# (PURCHASES and PURCHASES_FREQUENCY = negative),Keep high balances, mostly pay the minimum(BALANCE = 0.67, 
# MINIMUM_PAYMENTS = 0.53), and rarely clear their debit fully (PRC_FULL_PAYMENT = -0.45).
# These are credit-dependent or risky users — they rely on their card as a cash loan, not for purchases or borrowing 
# rather than shopping. They have high balances and low repayment discipline, meaning higher credit risk.


# Cluster 2 : Inactive / Low-Engagement Customers( 1,444 customers):
# Almost all features are negative like low BALANCE, PURCHASES, PAYMENTS, and FREQUENCIES.They have Very little spending
# or repayment activity, rarely update their balances, some pay in full, but activity is generally low(PRC_FULL_PAYMENT 
# = 0.22.)
# These are inactive or dormant customers — they rarely use their card and contribute little revenue.Banks might target 
# them with activation offers or reward campaigns to increase usage.


# Cluster 3 : Active Frequent Users with Moderate Spending(2,765 customers):
# They use card frequently, though not for huge amounts(0.80). They try keep it updated often(0.34) and are relatively
# responsible payers and their MINIMUM_PAYMENTS and PAYMENTS showing below average but not low.
# These are regular and balanced users and use their cards often but not heavily.They are medium-value, low-risk 
# customers who can be encouraged to spend more.


# radar (spider) chart

features = [
    'BALANCE', 'BALANCE_FREQUENCY', 'ONEOFF_PURCHASES', 'INSTALLMENTS_PURCHASES',
    'PURCHASES_FREQUENCY', 'ONEOFF_PURCHASES_FREQUENCY', 'CASH_ADVANCE_TRX',
    'PURCHASES_TRX', 'CREDIT_LIMIT', 'PAYMENTS', 'MINIMUM_PAYMENTS', 'PRC_FULL_PAYMENT'
]

cluster_means = df_final_with_id.groupby('Cluster')[features].mean()

cluster_means = (cluster_means - cluster_means.min()) / (cluster_means.max() - cluster_means.min())

angles = np.linspace(0, 2 * np.pi, len(features), endpoint=False)
angles = np.concatenate((angles, [angles[0]]))

plt.figure(figsize=(7,7), dpi=100)
ax = plt.subplot(111, polar=True)

for i, row in cluster_means.iterrows():
    values = np.concatenate((row.values, [row.values[0]]))
    ax.plot(angles, values, label=f'Cluster {i}')
    ax.fill(angles, values, alpha=0.1)

ax.set_xticks(angles[:-1])
ax.set_xticklabels(features, fontsize=8)
plt.title("Cluster Profiles (Radar Chart)")
plt.legend(loc='upper right', bbox_to_anchor=(1.2, 1.1))
plt.tight_layout()
plt.show()


# This spider chart visually summarizes how the four customer clusters differ in their average credit card behavior. Each
# color represents one cluster, showing how its members compare in spending, borrowing, and repayment patterns.
# Each axis (or spoke) corresponds to a feature that reflects customer activity, such as balance (average amount owed),
# purchases (total spending), cash advance transactions (use of the card for cash withdrawals), payments (total repayments
# ), and the percentage of full payments (how often customers pay off their balance completely). The distance from the
# center shows the strength of each behavior. Points near the center indicate lower activity, while points farther out 
# represent higher usage or stronger performance. All values are scaled between 0 and 1 to make features comparable.

# In the chart, the blue area(cluster 0) stretches widely across most axes, showing customers with high spending, frequent purchases,
# and strong repayment habits. The orange areas(cluster 1) peaks sharply on cash advances, representing customers who 
# rely on their credit cards mainly for cash and often carry debt. The green line area(cluster 2) close to the center, 
# indicating inactive or low-usage accounts with minimal transactions. The red area( cluster 3) shows moderate spread
# across most features, describing regular and balanced users who maintain steady but controlled credit activity.


from sklearn.decomposition import PCA

numeric_cols = df_final.select_dtypes(include=['number']).columns
df_numeric = df_final[numeric_cols]

#  PCA to 2 components
pca = PCA(n_components=2, random_state=42)
pca_components = pca.fit_transform(df_numeric)

pca_df = pd.DataFrame(data=pca_components, columns=['PC1', 'PC2'])
pca_df['Cluster'] = df_final_with_id['Cluster']


plt.figure(figsize=(10, 7))
sns.scatterplot(
    data=pca_df,
    x='PC1', y='PC2',
    hue='Cluster',
    palette='Set2',
    s=60,
    alpha=0.7
)
plt.title('Customer Clusters (PCA 2D Projection)', fontsize=14)
plt.xlabel('Principal Component 1')
plt.ylabel('Principal Component 2')
plt.legend(title='Cluster', loc='best')
plt.grid(True)
plt.show()

print("Explained variance ratio:", pca.explained_variance_ratio_.round(3))

Explained variance ratio: [0.308 0.246]


# This 2D PCA plot shows how customers are grouped into four distinct clusters based on their overall credit card behavior
# after reducing the original features to two principal components(PCA)for visualization. Each color represents a 
# different cluster, and the position of the points shows how similar customers are in terms of their usage patterns.
# It is like points that are close together share similar financial behaviors.
# The orange cluster (Cluster 0) occupies the top-left region.The pink cluster (Cluster 1) lies just below it.The green 
# cluster (Cluster 2) dominates the central region and the blue cluster (Cluster 3) is positioned toward the lower-left.
# The axes, “Principal Component 1” and “Principal Component 2,” are combinations of all original features, capturing the 
# two largest sources of variation in the data. The explained variance ratio (about 30.8% and 24.6%) shows that these two
# components together describe roughly 55% of the total behavioral variation among customers, allowing clear visualization
# of how groups differ.


import plotly.express as px


# PCA with 3 components
pca_3d = PCA(n_components=3, random_state=42)
pca_3d_components = pca_3d.fit_transform(df_numeric)

pca_3d_df = pd.DataFrame(pca_3d_components, columns=['PC1', 'PC2', 'PC3'])
pca_3d_df['Cluster'] = df_final_with_id['Cluster']
pca_3d_df['CUST_ID'] = df_final_with_id['CUST_ID']

hover_features = ['BALANCE', 'PURCHASES', 'CREDIT_LIMIT', 'PAYMENTS']
existing_hover_features = [col for col in hover_features if col in df_final_with_id.columns]

for col in existing_hover_features:
    pca_3d_df[col] = df_final_with_id[col]

fig = px.scatter_3d(
    pca_3d_df,
    x='PC1', y='PC2', z='PC3',
    color='Cluster',
    hover_data=['CUST_ID'] + existing_hover_features,
    title='Customer Clusters (PCA 3D Projection with Hover Info)',
    color_continuous_scale='Viridis',
    opacity=0.8
)


fig.update_traces(marker=dict(size=5))
fig.update_layout(
    scene=dict(
        xaxis_title='Principal Component 1',
        yaxis_title='Principal Component 2',
        zaxis_title='Principal Component 3'
    ),
    legend_title='Cluster'
)

fig.show()


print("Explained variance ratio:", pca_3d.explained_variance_ratio_.round(3))
print("Total variance explained:", round(sum(pca_3d.explained_variance_ratio_), 3))

Explained variance ratio: [0.308 0.246 0.095]
Total variance explained: 0.649


loadings = pd.DataFrame(
    pca_3d.components_.T,
    columns=['PC1', 'PC2', 'PC3'],
    index=df_numeric.columns
)
loadings.round(3)


# This table shows how each original feature contributes to the three principal components (PC1, PC2, and PC3) used in 
# the PCA analysis. Each number represents a loading, meaning the strength and direction of that feature’s influence on 
# each principal component.

# PC1 has strong positive loadings on PURCHASES_TRX, ONEOFF_PURCHASES, PURCHASES_FREQUENCY, and PAYMENTS which means PC1
# mainly represents overall purchasing and payment activity. Higher values indicate customers who make many purchases and
# payments.

# PC2 is most influenced by BALANCE, MINIMUM_PAYMENTS, and CASH_ADVANCE_TRX.
# This component reflects credit usage and debt behavior, capturing how much balance customers carry and whether they rely
# on cash advances or make only minimum payments.

# PC3 is strongly affected by CREDIT_LIMIT, PAYMENTS, and PRC_FULL_PAYMENT.
# This one highlights credit capacity and repayment responsibility and  higher scores indicate customers with larger 
# credit limits and more consistent full payments.

# so:

# PC1 = active spending and purchasing behavior
# PC2 = debt and balance management
# PC3 = repayment responsibility and credit capacity.


# tuning some parameters to see if there would be good K for better clustering. 

from sklearn.metrics import silhouette_score, davies_bouldin_score, calinski_harabasz_score

labels = kmeans.labels_
print("Silhouette:", silhouette_score(df_f, labels))
print("Davies-Bouldin:", davies_bouldin_score(df_f, labels))
print("Calinski-Harabasz:", calinski_harabasz_score(df_f, labels))

Silhouette: 0.21789570957959317
Davies-Bouldin: 1.5587120793545286
Calinski-Harabasz: 2062.029806851124


# Silhouette Score: Measures how well each point fits within its cluster compared to other clusters and ranges from –1 to
# +1 and +1 means points are very well grouped (clear separation).
# value (0.22) means the clusters are moderately distinct and not bad, but there’s some overlap.

# Davies–Bouldin Index: Measures the average similarity between clusters or how close and spread out they are. Lower is
# better which means small values mean clusters are compact and far apart.
# score (1.56) suggests some overlap between clusters they’re not perfectly separated.

# Calinski–Harabasz Index (2062.03): Compares cluster separation (between-cluster distance) to compactness 
# (within-cluster distance). Higher is better and large values mean dense, well-separated clusters.
# score (~2062) is quite solid, meaning the clusters are reasonably well defined overall.


from sklearn.metrics import silhouette_score

scores = []
for k in range(2, 10):
    kmeans = KMeans(n_clusters=k, random_state=42, n_init=20)
    labels = kmeans.fit_predict(df_f)
    score = silhouette_score(df_f, labels)
    scores.append((k, score))

pd.DataFrame(scores, columns=['k', 'Silhouette'])


# trying with K=2 and visulaizing it beside k=4

kmeans2 = KMeans(n_clusters=2, random_state=42, n_init=20)
kmeans4 = KMeans(n_clusters=4, random_state=42, n_init=20)

labels2 = kmeans2.fit_predict(df_numeric)
labels4 = kmeans4.fit_predict(df_numeric)

pca = PCA(n_components=2, random_state=42)
pca_components = pca.fit_transform(df_numeric)

pca_df = pd.DataFrame(pca_components, columns=["PC1", "PC2"])
pca_df["Cluster_k2"] = labels2
pca_df["Cluster_k4"] = labels4


fig, axes = plt.subplots(1, 2, figsize=(14, 6))

sns.scatterplot(
    data=pca_df, x="PC1", y="PC2", hue="Cluster_k2",
    palette="Set2", s=60, alpha=0.7, ax=axes[0]
)
axes[0].set_title("K-Means with k = 2", fontsize=13)
axes[0].set_xlabel("Principal Component 1")
axes[0].set_ylabel("Principal Component 2")
axes[0].legend(title="Cluster")

sns.scatterplot(
    data=pca_df, x="PC1", y="PC2", hue="Cluster_k4",
    palette="Set1", s=60, alpha=0.7, ax=axes[1]
)
axes[1].set_title("K-Means with k = 4", fontsize=13)
axes[1].set_xlabel("Principal Component 1")
axes[1].set_ylabel("Principal Component 2")
axes[1].legend(title="Cluster")

plt.suptitle("Comparison of Clustering Structures (k=2 vs k=4)", fontsize=15)
plt.tight_layout()
plt.show()


labels = kmeans2.labels_
print("Silhouette:", silhouette_score(df_f, labels))
print("Davies-Bouldin:", davies_bouldin_score(df_f, labels))
print("Calinski-Harabasz:", calinski_harabasz_score(df_f, labels))

Silhouette: 0.24626855324384822
Davies-Bouldin: 1.7593020764968328
Calinski-Harabasz: 2120.608247594492


In Silhouette higher (k=2),the two big groups are more separated, because there are fewer cluster boundaries.
For Davies–Bouldin lower (k=4), the four smaller clusters are more compact and balanced internally.
For Calinski–Harabasz, both( k=2, k=4) are strong, meaning the data has real structure either way.


While k = 2 performs slightly better numerically (with a higher Silhouette score), it only divides customers into two
broad, general categories and misses finer distinctions.
In contrast, k = 4 produces more detailed and interpretable segments that better reflect real-world customer diversity.
It identifies meaningful behavioral types, such as high spenders, cash users, low-activity customers, and balanced users.
Therefore, k = 4 provides a richer and more actionable understanding of customer behavior, even if the metrics are only
slightly lower.


# Now trying GMM methods (Gaussian Mixture Model)


from sklearn.mixture import GaussianMixture
from sklearn.metrics import silhouette_score, davies_bouldin_score, calinski_harabasz_score

df_gmm = df_f.copy()


# Training GMM model with 2 components
gmm = GaussianMixture(n_components=2, random_state=42)
gmm_labels = gmm.fit_predict(df_gmm)

df_final_with_id['GMM_Cluster'] = gmm_labels


sil = silhouette_score(df_gmm, gmm_labels)
db = davies_bouldin_score(df_gmm, gmm_labels)
ch = calinski_harabasz_score(df_gmm, gmm_labels)

print(f"Silhouette Score: {sil:.4f}")
print(f"Davies-Bouldin Index: {db:.4f}")
print(f"Calinski-Harabasz Index: {ch:.4f}")

Silhouette Score: 0.1416
Davies-Bouldin Index: 2.5996
Calinski-Harabasz Index: 1089.9096


pca = PCA(n_components=2, random_state=42)
pca_components = pca.fit_transform(df_gmm)

pca_df = pd.DataFrame(pca_components, columns=['PC1', 'PC2'])
pca_df['Cluster'] = gmm_labels

# Plot
plt.figure(figsize=(8, 6))
sns.scatterplot(data=pca_df, x='PC1', y='PC2', hue='Cluster', palette='Set2', s=60, alpha=0.7)
plt.title("Gaussian Mixture Model (GMM) Clustering - 2D PCA Projection")
plt.xlabel("Principal Component 1")
plt.ylabel("Principal Component 2")
plt.legend(title="Cluster")
plt.grid(True)
plt.show()

print("Explained variance ratio:", pca.explained_variance_ratio_.round(3))

Explained variance ratio: [0.308 0.246]


cluster_summary = df_final_with_id.groupby('GMM_Cluster')[df_final_with_id.select_dtypes(include='number').columns].mean().round(2)

cluster_summary['Count'] = df_final_with_id['GMM_Cluster'].value_counts().sort_index()
cluster_summary


# trying to to change some key parameter(covariance_typ) of GMM

from sklearn.mixture import GaussianMixture
from sklearn.metrics import silhouette_score

cov_types = ['full', 'tied', 'diag', 'spherical']
best_score = -1
best_params = None

for cov in cov_types:
    gmm = GaussianMixture(n_components=2, covariance_type=cov, random_state=42, n_init=5, reg_covar=1e-4)
    labels = gmm.fit_predict(df_f)
    sil = silhouette_score(df_f, labels, sample_size=2000, random_state=42)
    print(f"Covariance: {cov}, Silhouette: {sil:.4f}")
    
    if sil > best_score:
        best_score = sil
        best_params = cov

print("\n Best covariance type:", best_params, "with Silhouette:", round(best_score, 4))

Covariance: full, Silhouette: 0.1292
Covariance: tied, Silhouette: 0.2623
Covariance: diag, Silhouette: 0.1369
Covariance: spherical, Silhouette: 0.1219

 Best covariance type: tied with Silhouette: 0.2623


# Cov type /tied' is 0.262 which is the Best separation , compact, stable clusters.


# Now trying the best version K=2, and Covariance: tied
best_gmm = GaussianMixture(n_components=2, covariance_type='tied', random_state=42, n_init=5, reg_covar=1e-4)
best_gmm_labels = best_gmm.fit_predict(df_f)

sil = silhouette_score(df_f, best_gmm_labels, sample_size=2000, random_state=42)
print("Final Silhouette:", round(sil, 4))

df_final_with_id['GMM_Cluster'] = best_gmm_labels

cluster_summary = (
    df_final_with_id.drop(columns=['CUST_ID'], errors='ignore')
    .groupby('GMM_Cluster')
    .mean()
    .round(2)
)
cluster_summary['Count'] = df_final_with_id['GMM_Cluster'].value_counts().sort_index()
cluster_summary

Final Silhouette: 0.2623


pca = PCA(n_components=4, random_state=42)
pca_result = pca.fit_transform(df_f)


pca_df = pd.DataFrame(pca_result, columns=['PC1', 'PC2', 'PC3', 'PC4'])
pca_df['Cluster'] = df_final_with_id['GMM_Cluster']

plt.figure(figsize=(9, 7))
sns.scatterplot(
    data=pca_df,
    x='PC1', y='PC2',
    hue='Cluster',
    palette='Set2',
    s=70, alpha=0.7
)
plt.title("GMM (k=2, covariance='tied') Clusters - 2D PCA Projection", fontsize=14)
plt.xlabel("Principal Component 1")
plt.ylabel("Principal Component 2")
plt.legend(title="Cluster")
plt.grid(True)
plt.show()


# K-Means and GMM both uncover meaningful customer groups but differ in how they define them. K-Means creates clear, 
# distinct clusters, while GMM models softer, overlapping boundaries between groups. Although GMM with n=2 gives a 
# slightly higher Silhouette score (0.26), it forms less sharply separated clusters. K-Means with k=4 provides still 
# cleaner, more interpretable segments, making it better for practical customer analysis.
# However, if the goal is a probabilistic approach for estimating the likelihood that each customer belongs to multiple 
# clusters—GMM is the preferred model.


# the probabality that every costumers belong to which cluster # k=2
probs = best_gmm.predict_proba(df_f)

max_probs = probs.max(axis=1)  

conf_df = pd.DataFrame({
    'CUST_ID': df_final_with_id['CUST_ID'],
    'Confidence': max_probs
}).set_index('CUST_ID').sort_values('Confidence', ascending=False)

plt.figure(figsize=(6, 8))
sns.heatmap(
    conf_df[['Confidence']],
    cmap='coolwarm',
    cbar=True,
    linewidths=0.3
)
plt.title("Customer Membership Probabilities by Cluster (GMM)", fontsize=13)
plt.xlabel("Confidence")
plt.ylabel("Customer ID")
plt.show()


This graph shows how confident or the probabality the GMM model is about assigning each customer to their cluster.Each 
row represents one customer, and the color indicates their membership confidence. Dark red means the model is very sure
the customer fits their cluster (high probability, close to 1). Lighter or bluish shades mean the model is less certain,
suggesting the customers behavior overlaps between clusters. For example, if Customer C11300 appears in dark red, it
means the model is highly confident that this customer clearly belongs to their assigned cluster.


# In the next we will try hierarchical clustering


from sklearn.cluster import AgglomerativeClustering
from scipy.cluster.hierarchy import linkage, dendrogram


linkage_matrix = linkage(df_f, method='ward')  


plt.figure(figsize=(12, 5))
dendrogram(linkage_matrix, truncate_mode='level', p=5) 
plt.title("Hierarchical Clustering Dendrogram (truncated)")
plt.xlabel("Sample Index or (Cluster Size)")
plt.ylabel("Distance")
plt.show()


# This dendrogram shows how customer groups were formed step by step in hierarchical clustering. Each branch is showing 
# clusters being merged together based on how similar they are. The y-axis shows the distance or dissimilarity between
# clusters when they merge. If we cut the tree at distance = 100, we get about 4 clusters, while cutting at 125 merges 
# everything into 1 cluster. It means the lower branches show smaller, similar groups, and the higher ones show broader,
# combined clusters.


agg = AgglomerativeClustering(
    n_clusters=4,          
    metric='euclidean',    
    linkage='ward'         
)

df_final_with_id['Agglo_Cluster'] = agg.fit_predict(df_f)

print(df_final_with_id['Agglo_Cluster'].value_counts())

Agglo_Cluster
2    3326
0    2281
3    1684
1    1659
Name: count, dtype: int64


labels = df_final_with_id['Agglo_Cluster']

silhouette = silhouette_score(df_f, labels)
dbi = davies_bouldin_score(df_f, labels)
ch = calinski_harabasz_score(df_f, labels)

# Display results
print(f"Silhouette Score: {silhouette:.4f}")
print(f"Davies-Bouldin Index: {dbi:.4f}")
print(f"Calinski-Harabasz Index: {ch:.4f}")

Silhouette Score: 0.1344
Davies-Bouldin Index: 1.7837
Calinski-Harabasz Index: 1602.7016


pca = PCA(n_components=2, random_state=42)
pca_components = pca.fit_transform(df_f)

pca_df = pd.DataFrame(pca_components, columns=['PC1', 'PC2'])
pca_df['Cluster'] = df_final_with_id['Agglo_Cluster']

plt.figure(figsize=(8,6))
sns.scatterplot(
    data=pca_df,
    x='PC1', y='PC2',
    hue='Cluster',
    palette='Set2',
    s=60
)
plt.title("Agglomerative (Hierarchical) Clustering - 2D PCA Projection")
plt.xlabel("Principal Component 1")
plt.ylabel("Principal Component 2")
plt.legend(title='Cluster')
plt.grid(True, linestyle='--', alpha=0.5)
plt.show()

print("Explained variance ratio (PC1 + PC2):", round(sum(pca.explained_variance_ratio_), 3))

Explained variance ratio (PC1 + PC2): 0.554


cluster_summary = (
    df_final_with_id
    .groupby('Agglo_Cluster')[df_final_with_id.select_dtypes(include='number').columns]
    .mean()
    .round(2)
)

cluster_summary['Count'] = df_final_with_id['Agglo_Cluster'].value_counts().sort_index()
cluster_summary


# doing some parameter tuning


# different n_clusters

for k in range(2, 9):
    agg = AgglomerativeClustering(n_clusters=k, metric='euclidean', linkage='ward')
    labels = agg.fit_predict(df_f)
    sil = silhouette_score(df_f, labels)
    print(f"k={k}, Silhouette={sil:.4f}")

k=2, Silhouette=0.2501
k=3, Silhouette=0.2062
k=4, Silhouette=0.1344
k=5, Silhouette=0.1410
k=6, Silhouette=0.1501
k=7, Silhouette=0.1325
k=8, Silhouette=0.1388


# different linkage

for link in ['ward', 'complete', 'average', 'single']:
    agg = AgglomerativeClustering(n_clusters=4, metric='euclidean', linkage=link)
    labels = agg.fit_predict(df_f)
    sil = silhouette_score(df_f, labels)
    print(f"Linkage={link}, Silhouette={sil:.4f}")

Linkage=ward, Silhouette=0.1344
Linkage=complete, Silhouette=0.5470
Linkage=average, Silhouette=0.7155
Linkage=single, Silhouette=0.8121


# different affinity or metrics

for metric in ['euclidean', 'manhattan', 'cosine']:
    agg = AgglomerativeClustering(n_clusters=4, metric=metric, linkage='average')
    labels = agg.fit_predict(df_f)
    sil = silhouette_score(df_f, labels)
    print(f"Metric={metric}, Silhouette={sil:.4f}")

Metric=euclidean, Silhouette=0.7155
Metric=manhattan, Silhouette=0.7285
Metric=cosine, Silhouette=0.1030


#  Fitting the tuned model
agg_best = AgglomerativeClustering(
    n_clusters=2,           # best number of clusters
    metric='manhattan',      # best distance metric
    linkage='average'       # best linkage method
)

df_final_with_id['Agglo_Best'] = agg_best.fit_predict(df_f)

labels = df_final_with_id['Agglo_Best']

silhouette = silhouette_score(df_f, labels)
dbi = davies_bouldin_score(df_f, labels)
ch = calinski_harabasz_score(df_f, labels)

print(f"Tuned Agglomerative Model Results:")
print(f"Silhouette Score: {silhouette:.4f}")
print(f"Davies-Bouldin Index: {dbi:.4f}")
print(f"Calinski-Harabasz Index: {ch:.4f}")

Tuned Agglomerative Model Results:
Silhouette Score: 0.8223
Davies-Bouldin Index: 0.1254
Calinski-Harabasz Index: 52.9758


df_final_with_id['Agglo_Best'].value_counts()

Agglo_Best
0    8949
1       1
Name: count, dtype: int64


# The Agglomerative model show excellent metric values (Silhouette 0.82, Davies–Bouldin 0.12), but they are misleading.
# Almost all customers (8,949) fall into one cluster, and only one customer forms the second cluster.
# This extreme imbalance means the model didn’t find real structure and we can say its “good” scores come from poor 
# cluster distribution in which it could not classify the financial bhevavior of the customers in a good way. 
# In contrast, K-Means (k = 4) created balanced, interpretable clusters with realistic separation (Silhouette ≈ 0.22).
# GMM gave similar results but with overlapping, probabilistic clusters.
# While Agglomerative seems perfect by numbers, it’s not meaningful for analysis or segmentation.Therefore, K-Means 
# (k = 4) remains still the best and most practical clustering model for your dataset.


# Now goign for DBscan :

from sklearn.cluster import DBSCAN

dbscan = DBSCAN(eps=1.5, min_samples=5, metric='euclidean')
dbscan_labels = dbscan.fit_predict(df_f)

df_final_with_id['DBSCAN_Cluster'] = dbscan_labels

print(df_final_with_id['DBSCAN_Cluster'].value_counts().sort_index())

DBSCAN_Cluster
-1     616
 0    8324
 1      10
Name: count, dtype: int64


mask = df_final_with_id['DBSCAN_Cluster'] != -1  # Excluding noise
X_valid = df_f[mask]
labels_valid = df_final_with_id.loc[mask, 'DBSCAN_Cluster']

if len(set(labels_valid)) > 1:  # Avoid single cluster
    sil = silhouette_score(X_valid, labels_valid)
    db = davies_bouldin_score(X_valid, labels_valid)
    ch = calinski_harabasz_score(X_valid, labels_valid)

    print(f"Silhouette Score: {sil:.4f}")
    print(f"Davies-Bouldin Index: {db:.4f}")
    print(f"Calinski-Harabasz Index: {ch:.4f}")

Silhouette Score: 0.2009
Davies-Bouldin Index: 0.9895
Calinski-Harabasz Index: 18.1088


# Tuning Parameters (ε and min_samples)
# We’ll test multiple combinations to find the best.


results = []

for eps in np.arange(0.5, 3.0, 0.2):
    for min_samples in [3, 5, 8, 10]:
        model = DBSCAN(eps=eps, min_samples=min_samples, metric='euclidean')
        labels = model.fit_predict(df_f)

        # excluding noise (-1)
        unique_labels = set(labels)
        n_clusters = len(unique_labels - {-1})

        
        if n_clusters < 2:
            continue

       
        mask = labels != -1
        sil = silhouette_score(df_f[mask], labels[mask])
        results.append((eps, min_samples, n_clusters, sil))


tune_df = pd.DataFrame(results, columns=['eps', 'min_samples', 'n_clusters', 'Silhouette'])
print(tune_df.sort_values(by='Silhouette', ascending=False).head(10))

    eps  min_samples  n_clusters  Silhouette
28  2.7            3           2    0.680921
29  2.9            3           2    0.680452
25  1.9            5           2    0.510697
27  2.5            3           2    0.510079
26  2.1            3           3    0.466453
24  1.9            3          10    0.215052
20  1.5            5           2    0.200896
23  1.7            8           2    0.194679
22  1.7            5           3    0.191200
18  1.3            8           2    0.175377


# we will try eps = 2.7, min_samples = 3, giving 2 clusters with a Silhouette Score ≈ 0.68.


dbscan_best = DBSCAN(eps=2.7, min_samples=3, metric='euclidean')
dbscan_labels = dbscan_best.fit_predict(df_f)

df_final_with_id['DBSCAN_Best'] = dbscan_labels
print(df_final_with_id['DBSCAN_Best'].value_counts().sort_index())

DBSCAN_Best
-1      50
 0    8897
 1       3
Name: count, dtype: int64


mask = df_final_with_id['DBSCAN_Best'] != -1
X_valid = df_f[mask]
labels_valid = df_final_with_id.loc[mask, 'DBSCAN_Best']

sil = silhouette_score(X_valid, labels_valid)
dbi = davies_bouldin_score(X_valid, labels_valid)
ch = calinski_harabasz_score(X_valid, labels_valid)

print(f"Silhouette Score: {sil:.4f}")
print(f"Davies-Bouldin Index: {dbi:.4f}")
print(f"Calinski-Harabasz Index: {ch:.4f}")

Silhouette Score: 0.6809
Davies-Bouldin Index: 0.3398
Calinski-Harabasz Index: 50.7926


# DBSCAN also like Hierachial model mostly grouped almost all customers into one large cluster, with a few outliers and a 
# tiny secondary group.It produced high metric scores, but they’re misleading because the clustering is unbalanced and not
# meaningful. This happened maybe because DBSCAN relies on density differences, and the data is uniform and continuous, not 
# forming distinct dense regions. As a result, K-Means still performes better for this dataset, giving clearer and more 
# useful customer segments.


# Conclusion:

# In this study, several clustering algorithms like K-Means, Gaussian Mixture Model (GMM), Agglomerative (Hierarchical) 
# Clustering, and DBSCAN were applied to segment customers based on their credit card usage behavior. The goal was to
# identify meaningful groups of customers with similar spending, payment, and balance patterns.
# After data scaling and dimensionality reduction using PCA, we evaluated each model using performance metrics such as 
# the Silhouette Score, Davies–Bouldin Index, and Calinski–Harabasz Index. Although the Agglomerative and DBSCAN models 
# showed numerically high scores, both suffered from imbalanced results, producing one dominant cluster and a few outliers
# which makes them statistically good but practically meaningless.
# The GMM model performed slightly better in capturing probabilistic, overlapping clusters, but its separation was still
# weak.In contrast, K-Means with k = 4 provided balanced, interpretable, and realistic clusters, to some degree clearly
# distinguishing high spenders, cash users, low-activity customers, and moderate, responsible users.
# Overall, K-Means (k = 4) proved to be the most effective and actionable model for this dataset, offering valuable 
# insights into customer behavior that can support targeted marketing, credit management, and customer retention
# strategies.

	CUST_ID	BALANCE	BALANCE_FREQUENCY	PURCHASES	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	CASH_ADVANCE	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	PURCHASES_INSTALLMENTS_FREQUENCY	CASH_ADVANCE_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	TENURE
0	C10001	40.900749	0.818182	95.40	0.00	95.40	0.000000	0.166667	0.000000	0.083333	0.000000	0	2	1000.0	201.802084	139.509787	0.000000	12
1	C10002	3202.467416	0.909091	0.00	0.00	0.00	6442.945483	0.000000	0.000000	0.000000	0.250000	4	0	7000.0	4103.032597	1072.340217	0.222222	12
2	C10003	2495.148862	1.000000	773.17	773.17	0.00	0.000000	1.000000	1.000000	0.000000	0.000000	0	12	7500.0	622.066742	627.284787	0.000000	12
3	C10004	1666.670542	0.636364	1499.00	1499.00	0.00	205.788017	0.083333	0.083333	0.000000	0.083333	1	1	7500.0	0.000000	NaN	0.000000	12
4	C10005	817.714335	1.000000	16.00	16.00	0.00	0.000000	0.083333	0.083333	0.000000	0.000000	0	1	1200.0	678.334763	244.791237	0.000000	12
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
8945	C19186	28.493517	1.000000	291.12	0.00	291.12	0.000000	1.000000	0.000000	0.833333	0.000000	0	6	1000.0	325.594462	48.886365	0.500000	6
8946	C19187	19.183215	1.000000	300.00	0.00	300.00	0.000000	1.000000	0.000000	0.833333	0.000000	0	6	1000.0	275.861322	NaN	0.000000	6
8947	C19188	23.398673	0.833333	144.40	0.00	144.40	0.000000	0.833333	0.000000	0.666667	0.000000	0	5	1000.0	81.270775	82.418369	0.250000	6
8948	C19189	13.457564	0.833333	0.00	0.00	0.00	36.558778	0.000000	0.000000	0.000000	0.166667	2	0	500.0	52.549959	55.755628	0.250000	6
8949	C19190	372.708075	0.666667	1093.25	1093.25	0.00	127.040008	0.666667	0.666667	0.000000	0.333333	2	23	1200.0	63.165404	88.288956	0.000000	6

	BALANCE	BALANCE_FREQUENCY	PURCHASES	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	CASH_ADVANCE	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	PURCHASES_INSTALLMENTS_FREQUENCY	CASH_ADVANCE_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	TENURE
count	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8950.000000	8949.000000	8950.000000	8637.000000	8950.000000	8950.000000
mean	1564.474828	0.877271	1003.204834	592.437371	411.067645	978.871112	0.490351	0.202458	0.364437	0.135144	3.248827	14.709832	4494.449450	1733.143852	864.206542	0.153715	11.517318
std	2081.531879	0.236904	2136.634782	1659.887917	904.338115	2097.163877	0.401371	0.298336	0.397448	0.200121	6.824647	24.857649	3638.815725	2895.063757	2372.446607	0.292499	1.338331
min	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	50.000000	0.000000	0.019163	0.000000	6.000000
25%	128.281915	0.888889	39.635000	0.000000	0.000000	0.000000	0.083333	0.000000	0.000000	0.000000	0.000000	1.000000	1600.000000	383.276166	169.123707	0.000000	12.000000
50%	873.385231	1.000000	361.280000	38.000000	89.000000	0.000000	0.500000	0.083333	0.166667	0.000000	0.000000	7.000000	3000.000000	856.901546	312.343947	0.000000	12.000000
75%	2054.140036	1.000000	1110.130000	577.405000	468.637500	1113.821139	0.916667	0.300000	0.750000	0.222222	4.000000	17.000000	6500.000000	1901.134317	825.485459	0.142857	12.000000
max	19043.138560	1.000000	49039.570000	40761.250000	22500.000000	47137.211760	1.000000	1.000000	1.000000	1.500000	123.000000	358.000000	30000.000000	50721.483360	76406.207520	1.000000	12.000000

	CUST_ID	BALANCE	BALANCE_FREQUENCY	PURCHASES	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	CASH_ADVANCE	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	PURCHASES_INSTALLMENTS_FREQUENCY	CASH_ADVANCE_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	TENURE
0	C10001	40.900749	0.818182	95.40	0.00	95.40	0.000000	0.166667	0.000000	0.083333	0.000000	0	2	1000.0	201.802084	139.509787	0.000000	12
1	C10002	3202.467416	0.909091	0.00	0.00	0.00	6442.945483	0.000000	0.000000	0.000000	0.250000	4	0	7000.0	4103.032597	1072.340217	0.222222	12
2	C10003	2495.148862	1.000000	773.17	773.17	0.00	0.000000	1.000000	1.000000	0.000000	0.000000	0	12	7500.0	622.066742	627.284787	0.000000	12
3	C10004	1666.670542	0.636364	1499.00	1499.00	0.00	205.788017	0.083333	0.083333	0.000000	0.083333	1	1	7500.0	0.000000	312.343947	0.000000	12
4	C10005	817.714335	1.000000	16.00	16.00	0.00	0.000000	0.083333	0.083333	0.000000	0.000000	0	1	1200.0	678.334763	244.791237	0.000000	12
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
8945	C19186	28.493517	1.000000	291.12	0.00	291.12	0.000000	1.000000	0.000000	0.833333	0.000000	0	6	1000.0	325.594462	48.886365	0.500000	6
8946	C19187	19.183215	1.000000	300.00	0.00	300.00	0.000000	1.000000	0.000000	0.833333	0.000000	0	6	1000.0	275.861322	312.343947	0.000000	6
8947	C19188	23.398673	0.833333	144.40	0.00	144.40	0.000000	0.833333	0.000000	0.666667	0.000000	0	5	1000.0	81.270775	82.418369	0.250000	6
8948	C19189	13.457564	0.833333	0.00	0.00	0.00	36.558778	0.000000	0.000000	0.000000	0.166667	2	0	500.0	52.549959	55.755628	0.250000	6
8949	C19190	372.708075	0.666667	1093.25	1093.25	0.00	127.040008	0.666667	0.666667	0.000000	0.333333	2	23	1200.0	63.165404	88.288956	0.000000	6

	BALANCE	PURCHASES	CASH_ADVANCE	PAYMENTS	MINIMUM_PAYMENTS
0	3.735304	4.568506	0.000000	5.312231	4.945277
1	8.071989	0.000000	8.770896	8.319725	6.978531
2	7.822504	6.651791	0.000000	6.434654	6.442994
3	7.419183	7.313220	5.331694	0.000000	5.747301
4	6.707735	2.833213	0.000000	6.521114	5.504483

	BALANCE	PURCHASES	CASH_ADVANCE	PAYMENTS	MINIMUM_PAYMENTS
0	5.524100	6.874084	0.000000	9.475916	4.373865
1	19.863823	0.000000	6.190043	21.433774	5.877234
2	18.665676	12.252611	0.000000	13.123324	5.496177
3	16.849318	14.393644	4.290456	0.000000	4.985453
4	13.977587	3.629361	0.000000	13.439781	4.802909

	CASH_ADVANCE_TRX	PURCHASES_TRX
0	0.0	1.414214
1	2.0	0.000000
2	0.0	3.464102
3	1.0	1.000000
4	0.0	1.000000

	Feature	Skew_Before	Skew_Log	Skew_BoxCox	Kurt_Before	Kurt_Log	Kurt_BoxCox
BALANCE	BALANCE	2.39	-0.86	-0.12	7.67	0.17	-0.78
PURCHASES	PURCHASES	8.14	-0.76	-0.18	111.39	-0.82	-0.76
CASH_ADVANCE	CASH_ADVANCE	5.17	0.26	0.19	52.90	-1.73	-1.84
PAYMENTS	PAYMENTS	5.91	-1.78	0.12	54.77	6.06	1.33
MINIMUM_PAYMENTS	MINIMUM_PAYMENTS	13.85	0.27	-0.03	293.72	1.28	1.68

	BALANCE	BALANCE_FREQUENCY	PURCHASES	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	CASH_ADVANCE	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	PURCHASES_INSTALLMENTS_FREQUENCY	CASH_ADVANCE_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT
count	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00	8950.00
mean	0.00	0.00	-0.00	0.00	0.00	0.00	-0.00	-0.00	0.00	0.00	0.00	0.00	0.00	-0.00	-0.00	0.00
std	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00
min	-2.12	-3.70	-1.51	-0.36	-0.45	-0.94	-1.22	-0.68	-0.92	-0.68	-0.77	-1.14	-1.22	-2.78	-5.87	-0.53
25%	-0.82	0.05	-0.65	-0.36	-0.45	-0.94	-1.01	-0.68	-0.92	-0.68	-0.77	-0.74	-0.80	-0.61	-0.64	-0.53
50%	0.15	0.52	0.16	-0.33	-0.36	-0.94	0.02	-0.40	-0.50	-0.68	-0.77	-0.09	-0.41	-0.06	-0.12	-0.53
75%	0.72	0.52	0.71	-0.01	0.06	1.06	1.06	0.33	0.97	0.44	0.63	0.49	0.55	0.57	0.67	-0.04
max	2.73	0.52	3.56	24.20	24.43	1.73	1.27	2.67	1.60	6.82	7.00	6.33	7.01	4.57	4.04	2.89

	CUST_ID	BALANCE	BALANCE_FREQUENCY	PURCHASES	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	CASH_ADVANCE	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	PURCHASES_INSTALLMENTS_FREQUENCY	CASH_ADVANCE_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT
0	C10001	-1.238103	-0.249434	-0.368050	-0.356934	-0.349079	-0.944538	-0.806490	-0.678661	-0.707313	-0.675349	-0.772904	-0.578392	-0.960378	-0.983876	-0.824388	-0.525551
1	C10002	1.051961	0.134325	-1.505149	-0.356934	-0.454576	1.399909	-1.221758	-0.678661	-0.916995	0.573963	0.629493	-1.136616	0.688678	1.285321	0.917393	0.234227
2	C10003	0.860616	0.518084	0.521657	0.108889	-0.454576	-0.944538	1.269843	2.673451	-0.916995	-0.675349	-0.772904	0.230748	0.826100	-0.291722	0.475906	-0.525551
3	C10004	0.570542	-1.016953	0.875823	0.546189	-0.454576	0.680451	-1.014125	-0.399319	-0.916995	-0.258913	-0.071706	-0.741892	0.826100	-2.782085	-0.115812	-0.525551
4	C10005	0.111925	0.518084	-0.904786	-0.347294	-0.454576	-0.944538	-1.014125	-0.399319	-0.916995	-0.675349	-0.772904	-0.741892	-0.905410	-0.231669	-0.327305	-0.525551
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
8945	C19186	-1.351366	0.518084	0.066579	-0.356934	-0.132643	-0.944538	1.269843	-0.678661	1.179833	-0.675349	-0.772904	-0.169743	-0.960378	-0.708925	-1.782947	1.183951
8946	C19187	-1.465213	0.518084	0.079480	-0.356934	-0.122823	-0.944538	1.269843	-0.678661	1.179833	-0.675349	-0.772904	-0.169743	-0.960378	-0.807239	-0.115812	-0.525551
8947	C19188	-1.409349	-0.185477	-0.216300	-0.356934	-0.294893	-0.944538	0.854576	-0.678661	0.760469	-0.675349	-0.772904	-0.253986	-0.960378	-1.437791	-1.300748	0.329200
8948	C19189	-1.558537	-0.185477	-1.505149	-0.356934	-0.454576	0.237936	-1.221758	-0.678661	-0.916995	0.157527	0.218740	-1.136616	-1.097800	-1.626657	-1.660766	0.329200
8949	C19190	-0.324310	-0.889033	0.702026	0.301732	-0.454576	0.562485	0.439310	1.556082	-0.916995	0.990398	0.218740	0.756414	-0.905410	-1.549088	-1.237898	-0.525551

	BALANCE	BALANCE_FREQUENCY	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	Cluster	Count
Cluster
0	0.65	0.45	1.29	1.10	1.10	1.64	-0.08	1.60	1.02	1.07	0.49	0.24	0.0	1254
1	0.67	0.37	-0.26	-0.37	-0.81	-0.45	0.70	-0.71	-0.01	0.08	0.53	-0.45	1.0	3487
2	-1.34	-1.94	-0.25	-0.31	-0.52	-0.45	-0.50	-0.47	-0.24	-0.69	-1.07	0.22	2.0	1444
3	-0.45	0.34	-0.13	0.13	0.80	0.05	-0.58	0.41	-0.32	-0.22	-0.34	0.35	3.0	2765

	PC1	PC2	PC3
BALANCE	0.233	0.490	-0.094
BALANCE_FREQUENCY	0.238	0.240	-0.479
ONEOFF_PURCHASES	0.340	-0.069	0.267
INSTALLMENTS_PURCHASES	0.324	-0.119	0.003
PURCHASES_FREQUENCY	0.336	-0.291	-0.359
ONEOFF_PURCHASES_FREQUENCY	0.362	-0.133	-0.016
CASH_ADVANCE_TRX	-0.005	0.426	0.200
PURCHASES_TRX	0.425	-0.232	-0.202
CREDIT_LIMIT	0.280	0.115	0.485
PAYMENTS	0.343	0.145	0.372
MINIMUM_PAYMENTS	0.198	0.436	-0.156
PRC_FULL_PAYMENT	0.087	-0.348	0.288

	k	Silhouette
0	2	0.246269
1	3	0.230238
2	4	0.217896
3	5	0.204012
4	6	0.198933
5	7	0.197685
6	8	0.200939
7	9	0.202598

	BALANCE	BALANCE_FREQUENCY	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	Cluster	GMM_Cluster	Count
GMM_Cluster
0	0.52	0.52	-0.16	-0.19	-0.16	-0.09	0.27	-0.18	-0.04	-0.02	0.42	-0.48	1.49	0.0	4737
1	-0.58	-0.58	0.18	0.22	0.18	0.11	-0.31	0.20	0.04	0.02	-0.48	0.54	1.81	1.0	4213

	BALANCE	BALANCE_FREQUENCY	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	Cluster	Count
GMM_Cluster
0	-0.05	-0.08	-0.22	-0.10	-0.21	-0.41	0.05	-0.26	-0.12	-0.13	-0.02	-0.07	1.76	7395
1	0.25	0.40	1.04	0.47	0.99	1.94	-0.22	1.22	0.57	0.63	0.10	0.35	1.06	1555

	BALANCE	BALANCE_FREQUENCY	ONEOFF_PURCHASES	INSTALLMENTS_PURCHASES	PURCHASES_FREQUENCY	ONEOFF_PURCHASES_FREQUENCY	CASH_ADVANCE_TRX	PURCHASES_TRX	CREDIT_LIMIT	PAYMENTS	MINIMUM_PAYMENTS	PRC_FULL_PAYMENT	Cluster	GMM_Cluster	Agglo_Cluster	Count
Agglo_Cluster
0	-1.18	-1.12	-0.28	-0.11	0.07	-0.52	-0.55	-0.16	-0.37	-0.56	-0.92	0.59	2.35	0.01	0.0	2281
1	0.16	0.40	1.00	0.66	1.01	1.68	-0.43	1.26	0.55	0.65	0.02	0.44	1.18	0.80	1.0	1659
2	0.18	0.36	-0.21	-0.11	-0.22	-0.33	-0.09	-0.23	-0.33	-0.25	0.16	-0.41	1.76	0.03	2.0	3326
3	1.09	0.41	-0.20	-0.29	-0.66	-0.31	1.35	-0.56	0.61	0.61	0.91	-0.42	0.90	0.08	3.0	1684